一种改进的视网膜脱落图像分割算法研究

An Improved Algorithm for Segmentation of Retinal Detachment Images

为实现对超声图像病灶的分割并提高目标分割精度,提出一种残差结构的深层尺度融合的分割网络模型RDFA-Net。使用改进的网络模型并使用可变形卷积来增大采样范围,使用通道和空间注意力机制捕获特征,融入深层尺度特征进行融合以捕获深层次图像特征,成功将视网膜脱落病灶的超声图片进行分割,并有了更好的分割结果。使用合作医院提供的眼底超声图像作为数据自制分割数据集来评价RDFA-Net。通过实验结果对比显示,RDFA-Net是有效的,可以对视网膜脱落进行图像分割,并实现了较好的分割精度。

To achieve segmentation of lesions in ultrasound images and improve target segmentation accuracy,we propose a residual structured deep scale fusion attention network model,RDFA-Net.Using an improved residual structure based network model and deformable convolution to increase the sampling range,using channel and spatial attention mechanisms to capture features,and integrating deep scale features for fusion to capture deep level image features,the model successfully segmented the ultrasound images of retinal detachment lesions and produced better segmentation results.The self-made segmented dataset provided by the cooperative hospital is used to evaluate RDFA-Net.Through the comparison of experimental results,we found that RDFA-Net is effective and achieves advanced performance.